AI in Growth Marketing: From Prompt to Pipeline

Give the model a specific job

AI can help draft variants, organize research and prepare analysis. Useful output starts with the audience, the task and the evidence available. A prompt cannot supply a missing customer result or establish that an ad will work.

A brief that produces reviewable work

Separate drafting from launch

Ask for a small set of variants tied to distinct hypotheses. Review the claims, destination and tracking before publishing. When an agent has access to tools, define which actions require approval and verify the result after an authorized change.

A model’s confidence score is not a substitute for checking its work. Do not let a generated recommendation change budgets, send messages or launch a campaign without the approval and controls agreed for that workflow.

What a messaging experiment can establish

In Anthony’s prior work at Lyft, a July 2024 experiment compared rotating AI-assisted push copy, a static message and a no-message control. The reported rotating-copy result was +0.44% driver hours versus the no-message control. The separate comparisons do not establish a statistically significant rotating-versus-static difference or an AI-versus-human advantage. The underlying platform data has not been independently re-audited here.

Measure the workflow and the business separately

For the workflow, record time spent, review effort, errors and whether the output was usable. For the business, define the outcome, baseline, comparison and observation window. Faster drafting does not establish a conversion or revenue improvement.

Start with one bounded job and enough evidence to judge whether it helped. Agree the next step from the results.